MiniT2I: Pixel-Space Text-to-Image with MM-JiT at 258M Parameters

TL;DR

MiniT2I is a 258M-parameter text-to-image model from Kaiming He’s MIT group, authored by undergraduate Xianbang Wang. It is the first T2I extension of the JiT (“Back to Basics”) pixel-space diffusion line, built on a new MM-JiT architecture that fuses text and image streams without a VAE, tokenizer, or external feature alignment. The pitch is extreme parameter and compute frugality: the full T2I training run is reported to cost about the same as a single standard ImageNet experiment. The blog post (a hash-routed SPA at peppaking8.github.io) was not directly fetchable at filing; the details below are reconstructed from Chinese press coverage of the release and from the upstream JiT and PixelDiT papers.

Key claims

MiniT2I generates text-conditioned images directly in pixel space at 258M parameters total, without a VAE encoder/decoder, learned tokenizer, or external feature alignment (CLIP/DINO) [blog body — not directly verified; per 163.com coverage of the post].
The architecture is MM-JiT, a multi-modal extension of Back to Basics: Let Denoising Generative Models Denoise‘s JiT to T2I; JiT itself is “Just image Transformers” — a plain ViT that ingests raw pixels in large patches (up to 3072-d) and predicts clean image patches rather than noise [§Method of upstream JiT, summarized in 163.com coverage].
Total training cost for MiniT2I is reported as comparable to a single ImageNet diffusion experiment, in contrast to standard billion-parameter T2I models trained on massive image-text corpora [blog claim per 163.com summary].
Author cohort is notable: per the press release the project team is composed entirely of undergraduates from Kaiming He’s group [163.com coverage].

Method

MM-JiT is the T2I-conditioning extension of the JiT architecture introduced in Back to Basics: Let Denoising Generative Models Denoise. JiT’s central design choice is to abandon the VAE entirely and run a plain transformer on raw pixel patches, with the model predicting $\hat{x}_0$ (clean patches) rather than noise — this is the “x-prediction” choice that makes high-dimensional pixel diffusion tractable without a tokenizer.

MiniT2I extends this to text-conditioned generation by adding text-stream conditioning to the JiT backbone, in the same spirit as how MM-DiT extends DiT in PixelDiT: Pixel Diffusion Transformers for Image Generation‘s T2I variant (which also bolts MM-DiT-style fused text+image blocks onto a pixel-space backbone). The blog post is the only public artifact at filing time; no preprint or code release accompanies it as of 2026-06-18.

The training-cost claim — “one ImageNet experiment” — implies a budget on the order of what a 256² ImageNet class-conditional JiT run consumes, which is small in T2I terms but is not directly comparable to FLUX-class budgets without knowing the exact image-text dataset, resolution, and step count. The blog post is the source of the claim; the underlying numbers are not independently reported.

Results

The original blog post lists generated samples but no headline metrics were retrievable at filing time. Per secondary coverage:

258M parameters total, in contrast to typical T2I models in the 1–12B range.
Training cost “equivalent to a standard ImageNet experiment” (163.com summary).
Sample quality is described as “decent” rather than SOTA; the contribution is the parameter/compute efficiency at the pixel-space architecture choice, not a new benchmark high.

For reference, contemporaneous pixel-space T2I numbers: PixelDiT: Pixel Diffusion Transformers for Image Generation‘s PixelDiT-T2I at 1.3B params reports GenEval 0.74 / DPG 83.5 natively at 1024² (vs FLUX-dev’s 0.67/84.0 at 12B). MiniT2I sits well below this on parameters but no head-to-head GenEval/DPG numbers were available at filing.

Why it’s interesting

MiniT2I is the smallest filed instance of the “pixel-space diffusion can scale to T2I” thesis that this wiki has been tracking through Pixel-space diffusion. It sits at the extreme low-parameter, low-compute end of a recipe family that also includes PixelDiT: Pixel Diffusion Transformers for Image Generation (NVIDIA, 1.3B params, native 1024²) and Asymmetric Flow Models (Stanford, JiT-H/16 at 953M with rank-asymmetric velocity). All three share the JiT lineage from Back to Basics: Let Denoising Generative Models Denoise, and MiniT2I is the first to push the “no VAE, no tokenizer, no alignment” recipe to T2I at undergraduate-budget scale.

The training-cost claim — one ImageNet experiment for a T2I model — sharpens the Diffusion training efficiency story: if pixel-space + x-prediction + MM-DiT-style text fusion really gets you a usable T2I model at this budget, the choice to train modern T2I models at FLUX-class compute becomes a quality-vs-cost tradeoff rather than a necessity. The wiki currently has PRX Part 3 — Training a Text-to-Image Model in 24h as the closest comparable budget-T2I datapoint (24-hour latent T2I training) and MiniT2I would be its pixel-space analogue.

The blog is also a useful pointer to where the JiT line is heading in 2026: the MIT group’s primary research push, beyond pure ImageNet benchmarks, into actually conditional generation.